## TL;DR

Upgrade: memory accounting and usage were improved in 8.140.13, and 8.181.15 cut static memory of multivalue attributes by up to 40 percent. There is a fixed memory cost per document type and per field (the reporter had 39 schemas), so high startup memory with zero documents is expected on older 8.x versions.

## The problem

Vespa proton takes 20GB upon startup with zero documents deployed across many schemas (the reporter had 39).

## Fix it

1. Check your Vespa version.
   Success check: you are on an 8.x older than 8.140.13, confirming the fix applies.

2. Upgrade to 8.140.13 or newer (8.181.15+ preferred for the multivalue attribute savings).
   Success check: startup memory drops noticeably with the same schemas.

3. Inspect per-document-type usage via the searchnode metrics or the proton state API to confirm which types cost the most.
   Success check: metrics show the fixed per-type cost.

## When this applies

- Proton uses a lot of memory at startup with zero or few documents.
- Many schemas/document types deployed (dozens).
- You are on Vespa 8.x older than 8.140.13.

## When it does not apply

- Memory grows during feeding or querying; that is a different tuning problem.
- You are already on 8.181.15+ and memory is still high; profile with the state API.
- Vespa 7.x; the fixes are 8.x-specific.

## Compatibility

- Vespa 8.140.13+: improved memory accounting. 8.181.15+: multivalue attribute static memory cut up to 40 percent.

## Variant phrasings

- Vespa proton high memory startup zero documents
- proton 20GB memory startup
- Vespa memory per document type field fixed cost

## Why it happens

Proton allocates a fixed memory footprint per document type and per field at startup, before any documents arrive. With many schemas this adds up. Older 8.x versions also over-accounted multivalue attribute memory, fixed in the versions above.

## Edge cases

- Reducing schema count or consolidating fields lowers the fixed cost directly.
- Source: https://github.com/vespa-engine/vespa/issues/26350

---

## Vectle citations

- Exact version: <https://vectle.com/skills/skl_FLAx0g-1DO5V6f-8MxO4gg/versions/skv_9AsF_JDHOlvcPPYu_PxFuA>
- Canonical HTML skill page: <https://vectle.com/skills/skl_FLAx0g-1DO5V6f-8MxO4gg>
- Current Markdown URL: <https://vectle.com/skills/skl_FLAx0g-1DO5V6f-8MxO4gg/markdown>

## Selected public attribution and original notices/references

```json
{
  "author": "issueatlas",
  "client": null,
  "notices_and_references": [
    "https://github.com/vespa-engine/vespa/issues/26350"
  ]
}
```